Using machine learning-based Natural Language Processing to quantify emergency department presentations related to...

Suicide and self-harm are significant issues globally. Accurate, efficient and comprehensive data are required to identify people who present to Emergency Departments due to self-harm to receive current accepted interventions and to develop effective health policies and responses. Current methods for identifying people presenting with these behaviors can be time- and labor-intensive or can underestimate the true figure.

This study investigated the use of a novel machine learning-based Natural Language Processing program developed to quantify the number of Emergency Department presentations which were related to suicidal or self-harm ideation or behavior. The program identifies these presentations based on Emergency Department triage notes. We compared the Natural Language Processing program with alternative methods for identifying suicide or self-harm related presentations, including International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Australian Modification coding and keyword searching.

Using the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Australian Modification codes included with the dataset, 10,399 Emergency Department presentations related to suicide or self-harm were identified for the period July 2015 to June 2022, while the Natural Language Processing program found 27,298 presentations over the same period with a precision of 0.89 and a recall of 0.94. All methods were evaluated by comparing their identifications with a set of manually identified presentations. Natural Language Processing identification was the most appropriate for providing an accurate, comprehensive and efficient quantification.

This study revealed that less than 40% of Emergency Department presentations related to suicide or self-harm are identified using existing methods in the Australian Capital Territory. By providing an improved identification method, this study enables more accurate analysis and understanding of the issues of suicide and self-harm.